Skip to content

Introduction

AgentStateGraph is a content-addressed, versioned, branchable structured state store designed as an infrastructure primitive for intent-based systems.

AI agents don’t execute linear scripts — they explore state spaces. An agent asked to “set up a cluster for ML training” tries different approaches, compares outcomes, and picks winners. It needs to:

  • Branch to try approaches without risk
  • Compare outcomes side-by-side
  • Merge the winner back
  • Report what was done, what deviated, and why
  • Record the full reasoning chain for audit

No existing tool supports this natively. Git is text-oriented. Databases lack branching. Event sourcing is append-only.

Every state change in AgentStateGraph captures the full provenance chain:

FieldQuestion
state_rootWhat changed?
intentWhy?
reasoningHow did the agent decide?
confidenceHow sure was it?
agent_idWho did it?
authorityWho authorized it?
resolutionWhat was accomplished? Deviations?
  • Content-addressed Merkle DAG — immutable, deduplicated history
  • Schema-aware merge — CRDT-inspired conflict resolution
  • Speculative execution — O(1) branching, instant discard
  • Multi-agent orchestration — scoped sessions, delegation, intent trees
  • Epochs — sealable, tamper-evident audit bundles
  • 20 MCP tools — any agent can connect immediately